1; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s 2 3define <9 x double> @strided_load_3x3_volatile(double* %in, i64 %stride) { 4; CHECK-LABEL: @strided_load_3x3_volatile( 5; CHECK-NEXT: entry: 6; CHECK-NEXT: [[VEC_START:%.*]] = mul i64 0, [[STRIDE:%.*]] 7; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, double* %in, i64 [[VEC_START]] 8; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast double* [[VEC_GEP]] to <3 x double>* 9; CHECK-NEXT: load volatile <3 x double>, <3 x double>* [[VEC_CAST]], align 8 10; CHECK-NEXT: [[VEC_START1:%.*]] = mul i64 1, [[STRIDE]] 11; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, double* %in, i64 [[VEC_START1]] 12; CHECK-NEXT: [[VEC_CAST3:%.*]] = bitcast double* [[VEC_GEP2]] to <3 x double>* 13; CHECK-NEXT: load volatile <3 x double>, <3 x double>* [[VEC_CAST3]], align 8 14; CHECK-NEXT: [[VEC_START5:%.*]] = mul i64 2, [[STRIDE]] 15; CHECK-NEXT: [[VEC_GEP6:%.*]] = getelementptr double, double* %in, i64 [[VEC_START5]] 16; CHECK-NEXT: [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <3 x double>* 17; CHECK-NEXT: load volatile <3 x double>, <3 x double>* [[VEC_CAST7]], align 8 18; CHECK-NOT: = load 19; 20entry: 21 %load = call <9 x double> @llvm.matrix.column.major.load.v9f64(double* %in, i64 %stride, i1 true, i32 3, i32 3) 22 ret <9 x double> %load 23} 24 25declare <9 x double> @llvm.matrix.column.major.load.v9f64(double*, i64, i1, i32, i32) 26 27define <4 x double> @load_volatile_multiply(<4 x double>* %in) { 28; CHECK-LABEL: @load_volatile_multiply( 29; CHECK-NEXT: [[TMP1:%.*]] = bitcast <4 x double>* [[IN:%.*]] to double* 30; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast double* [[TMP1]] to <2 x double>* 31; CHECK-NEXT: load volatile <2 x double>, <2 x double>* [[VEC_CAST]], align 8 32; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP1]], i64 2 33; CHECK-NEXT: [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>* 34; CHECK-NEXT: load volatile <2 x double>, <2 x double>* [[VEC_CAST1]], align 8 35; CHECK-NOT: = load 36; 37 %in.m = load volatile <4 x double>, <4 x double>* %in, align 8 38 %res = call <4 x double> @llvm.matrix.multiply(<4 x double> %in.m, <4 x double> %in.m, i32 2, i32 2, i32 2) 39 ret <4 x double> %res 40} 41 42declare <4 x double> @llvm.matrix.multiply(<4 x double>, <4 x double>, i32, i32, i32) 43 44 45define <9 x double> @strided_load_3x3_align32(double* %in, i64 %stride) { 46; CHECK-LABEL: @strided_load_3x3_align32( 47; CHECK-NEXT: entry: 48; CHECK-NEXT: [[VEC_START:%.*]] = mul i64 0, [[STRIDE:%.*]] 49; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, double* %in, i64 [[VEC_START]] 50; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast double* [[VEC_GEP]] to <3 x double>* 51; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST]], align 32 52; CHECK-NEXT: [[VEC_START1:%.*]] = mul i64 1, [[STRIDE]] 53; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, double* %in, i64 [[VEC_START1]] 54; CHECK-NEXT: [[VEC_CAST3:%.*]] = bitcast double* [[VEC_GEP2]] to <3 x double>* 55; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST3]], align 8 56; CHECK-NEXT: [[VEC_START5:%.*]] = mul i64 2, [[STRIDE]] 57; CHECK-NEXT: [[VEC_GEP6:%.*]] = getelementptr double, double* %in, i64 [[VEC_START5]] 58; CHECK-NEXT: [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <3 x double>* 59; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST7]], align 8 60; CHECK-NOT: = load 61; 62entry: 63 %load = call <9 x double> @llvm.matrix.column.major.load.v9f64(double* align 32 %in, i64 %stride, i1 false, i32 3, i32 3) 64 ret <9 x double> %load 65} 66 67define <9 x double> @strided_load_3x3_align2(double* %in, i64 %stride) { 68; CHECK-LABEL: @strided_load_3x3_align2( 69; CHECK-NEXT: entry: 70; CHECK-NEXT: [[VEC_START:%.*]] = mul i64 0, [[STRIDE:%.*]] 71; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, double* %in, i64 [[VEC_START]] 72; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast double* [[VEC_GEP]] to <3 x double>* 73; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST]], align 2 74; CHECK-NEXT: [[VEC_START1:%.*]] = mul i64 1, [[STRIDE]] 75; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, double* %in, i64 [[VEC_START1]] 76; CHECK-NEXT: [[VEC_CAST3:%.*]] = bitcast double* [[VEC_GEP2]] to <3 x double>* 77; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST3]], align 2 78; CHECK-NEXT: [[VEC_START5:%.*]] = mul i64 2, [[STRIDE]] 79; CHECK-NEXT: [[VEC_GEP6:%.*]] = getelementptr double, double* %in, i64 [[VEC_START5]] 80; CHECK-NEXT: [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <3 x double>* 81; CHECK-NEXT: load <3 x double>, <3 x double>* [[VEC_CAST7]], align 2 82; CHECK-NOT: = load 83; 84entry: 85 %load = call <9 x double> @llvm.matrix.column.major.load.v9f64(double* align 2 %in, i64 %stride, i1 false, i32 3, i32 3) 86 ret <9 x double> %load 87} 88 89 90define <4 x double> @load_align2_multiply(<4 x double>* %in) { 91; CHECK-LABEL: @load_align2_multiply( 92; CHECK-NEXT: [[TMP1:%.*]] = bitcast <4 x double>* [[IN:%.*]] to double* 93; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast double* [[TMP1]] to <2 x double>* 94; CHECK-NEXT: load <2 x double>, <2 x double>* [[VEC_CAST]], align 2 95; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP1]], i64 2 96; CHECK-NEXT: [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>* 97; CHECK-NEXT: load <2 x double>, <2 x double>* [[VEC_CAST1]], align 2 98; CHECK-NOT: = load 99; 100 %in.m = load <4 x double>, <4 x double>* %in, align 2 101 %res = call <4 x double> @llvm.matrix.multiply(<4 x double> %in.m, <4 x double> %in.m, i32 2, i32 2, i32 2) 102 ret <4 x double> %res 103} 104 105define <6 x float> @strided_load_2x3_align16_stride2(float* %in) { 106; CHECK-LABEL: @strided_load_2x3_align16_stride2( 107; CHECK-NEXT: entry: 108; CHECK-NEXT: [[VEC_CAST:%.*]] = bitcast float* %in to <2 x float>* 109; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x float>, <2 x float>* [[VEC_CAST]], align 16 110; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr float, float* %in, i64 2 111; CHECK-NEXT: [[VEC_CAST1:%.*]] = bitcast float* [[VEC_GEP]] to <2 x float>* 112; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x float>, <2 x float>* [[VEC_CAST1]], align 8 113; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr float, float* %in, i64 4 114; CHECK-NEXT: [[VEC_CAST4:%.*]] = bitcast float* [[VEC_GEP3]] to <2 x float>* 115; CHECK-NEXT: [[COL_LOAD5:%.*]] = load <2 x float>, <2 x float>* [[VEC_CAST4]], align 16 116; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x float> [[COL_LOAD]], <2 x float> [[COL_LOAD2]], <4 x i32> <i32 0, i32 1, i32 2, i32 3> 117; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x float> [[COL_LOAD5]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef> 118; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP2]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5> 119; CHECK-NEXT: ret <6 x float> [[TMP3]] 120; 121entry: 122 %load = call <6 x float> @llvm.matrix.column.major.load.v6f32(float* align 16 %in, i64 2, i1 false, i32 2, i32 3) 123 ret <6 x float> %load 124} 125 126declare <6 x float> @llvm.matrix.column.major.load.v6f32(float*, i64, i1, i32, i32) 127